diff --git a/cmd/root.go b/cmd/root.go
index 3ac48f8d..59bbcd6f 100644
--- a/cmd/root.go
+++ b/cmd/root.go
@@ -164,6 +164,7 @@ func init() {
 	f.IntVar(&rootArgs.config.RateLimitRPS, "rate-limit-rps", defaultRateLimitRPS, "Maximum requests per second per IP for rate limiting")
 	f.IntVar(&rootArgs.config.RateLimitBurst, "rate-limit-burst", defaultRateLimitBurst, "Maximum burst size per IP for rate limiting")
 	f.BoolVar(&rootArgs.config.RateLimitFailClosed, "rate-limit-fail-closed", false, "On rate-limit backend errors, reject with 503 instead of allowing the request")
+	f.StringSliceVar(&rootArgs.config.TrustedProxies, "trusted-proxies", nil, "Comma-separated CIDRs of trusted reverse proxies. When set, gin uses X-Forwarded-For from these networks. Empty (default) trusts no proxies and uses RemoteAddr.")
 
 	// JWT flags
 	f.StringVar(&rootArgs.config.JWTType, "jwt-type", "", "Type of JWT to use")
@@ -472,6 +473,7 @@ func runRoot(c *cobra.Command, args []string) {
 	// Prepare server
 	deps := &server.Dependencies{
 		Log:          &log,
+		AppConfig:    &rootArgs.config,
 		HTTPProvider: httpProvider,
 	}
 	// Create the server
diff --git a/internal/config/config.go b/internal/config/config.go
index a43984e2..0d096ccf 100644
--- a/internal/config/config.go
+++ b/internal/config/config.go
@@ -259,4 +259,10 @@ type Config struct {
 	RateLimitBurst int
 	// RateLimitFailClosed rejects requests when the rate limit backend errors (default: fail-open).
 	RateLimitFailClosed bool
+	// TrustedProxies is the list of CIDRs allowed to set X-Forwarded-For
+	// and similar proxy headers. Empty (the default) means no proxies are
+	// trusted and gin will use RemoteAddr directly. Operators behind a
+	// reverse proxy MUST set this explicitly or rate limiting and audit
+	// logs will key on the proxy IP, not the real client IP.
+	TrustedProxies []string
 }
diff --git a/internal/rate_limit/in_memory.go b/internal/rate_limit/in_memory.go
index ca479626..dd5063b5 100644
--- a/internal/rate_limit/in_memory.go
+++ b/internal/rate_limit/in_memory.go
@@ -3,6 +3,7 @@ package rate_limit
 import (
 	"context"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"golang.org/x/time/rate"
@@ -13,9 +14,22 @@ import (
 // Compile-time interface guard
 var _ Provider = (*inMemoryProvider)(nil)
 
+// entry stores a per-IP rate limiter together with the last-seen timestamp
+// used by the cleanup goroutine. lastSeen is accessed concurrently by
+// Allow() and cleanup(), so it MUST be touched only via the atomic helpers
+// below — using a plain time.Time field would trip the race detector and
+// (more importantly) drop updates non-deterministically.
 type entry struct {
 	limiter  *rate.Limiter
-	lastSeen time.Time // benign data race: only used for cleanup staleness heuristic
+	lastSeen atomic.Int64 // unix nanoseconds
+}
+
+func (e *entry) touch() {
+	e.lastSeen.Store(time.Now().UnixNano())
+}
+
+func (e *entry) lastSeenTime() time.Time {
+	return time.Unix(0, e.lastSeen.Load())
 }
 
 type inMemoryProvider struct {
@@ -38,15 +52,11 @@ func newInMemoryProvider(cfg *config.Config, deps *Dependencies) (*inMemoryProvi
 
 // Allow checks if a request from the given IP is allowed
 func (p *inMemoryProvider) Allow(_ context.Context, ip string) (bool, error) {
-	v, loaded := p.visitors.LoadOrStore(ip, &entry{
-		limiter:  rate.NewLimiter(p.rps, p.burst),
-		lastSeen: time.Now(),
-	})
+	newEntry := &entry{limiter: rate.NewLimiter(p.rps, p.burst)}
+	newEntry.touch()
+	v, _ := p.visitors.LoadOrStore(ip, newEntry)
 	e := v.(*entry)
-	e.lastSeen = time.Now()
-	if loaded {
-		p.visitors.Store(ip, e)
-	}
+	e.touch()
 	return e.limiter.Allow(), nil
 }
 
@@ -67,7 +77,7 @@ func (p *inMemoryProvider) cleanup(ctx context.Context) {
 		case <-ticker.C:
 			p.visitors.Range(func(key, value any) bool {
 				e := value.(*entry)
-				if time.Since(e.lastSeen) > 10*time.Minute {
+				if time.Since(e.lastSeenTime()) > 10*time.Minute {
 					p.visitors.Delete(key)
 				}
 				return true
diff --git a/internal/rate_limit/redis.go b/internal/rate_limit/redis.go
index f1d27121..55f5cb2a 100644
--- a/internal/rate_limit/redis.go
+++ b/internal/rate_limit/redis.go
@@ -39,10 +39,15 @@ type redisProvider struct {
 }
 
 func newRedisProvider(cfg *config.Config, deps *Dependencies) (*redisProvider, error) {
-	// Window = burst / rps, minimum 1 second
+	// Window in seconds = ceil(burst / rps), minimum 1. The previous integer
+	// division silently truncated to 0 when burst < rps and produced
+	// inconsistent enforcement vs the in-memory limiter (which uses the
+	// same effective window via golang.org/x/time/rate). Use ceiling
+	// arithmetic so the redis window is at least as long as the rps period.
 	window := 1
 	if cfg.RateLimitRPS > 0 {
-		w := int(cfg.RateLimitBurst / cfg.RateLimitRPS)
+		// ceil(burst / rps) without floats: (a + b - 1) / b
+		w := (cfg.RateLimitBurst + cfg.RateLimitRPS - 1) / cfg.RateLimitRPS
 		if w > 1 {
 			window = w
 		}
@@ -55,13 +60,21 @@ func newRedisProvider(cfg *config.Config, deps *Dependencies) (*redisProvider, e
 	}, nil
 }
 
-// Allow checks if a request from the given IP is allowed using Redis
+// Allow checks if a request from the given IP is allowed using Redis.
+//
+// Errors are PROPAGATED to the caller (the rate-limit middleware) so that
+// the operator-controlled RateLimitFailClosed config can actually take
+// effect. Previously this returned (true, nil) on any redis error, which
+// meant fail-closed mode was a no-op and a flapping redis silently disabled
+// rate limiting entirely.
 func (p *redisProvider) Allow(ctx context.Context, ip string) (bool, error) {
 	key := "rate_limit:" + ip
 	result, err := p.client.Eval(ctx, rateLimitScript, []string{key}, p.burst, p.window).Int64()
 	if err != nil {
-		p.log.Error().Err(err).Str("ip", ip).Msg("rate limit redis error, failing open")
-		return true, nil
+		p.log.Error().Err(err).Str("ip", ip).Msg("rate limit redis error")
+		// Default to allowing the request, but PROPAGATE the error so the
+		// middleware can fail-closed when configured.
+		return true, err
 	}
 	return result == 1, nil
 }
diff --git a/internal/server/http_routes.go b/internal/server/http_routes.go
index ca332f9d..63d4bd11 100644
--- a/internal/server/http_routes.go
+++ b/internal/server/http_routes.go
@@ -11,6 +11,17 @@ import (
 // NewRouter creates new gin router
 func (s *server) NewRouter() *gin.Engine {
 	router := gin.New()
+	// Restrict the set of proxies whose forwarded headers are honoured.
+	// When TrustedProxies is empty/nil, gin trusts NO proxies and falls back
+	// to RemoteAddr — preventing X-Forwarded-For spoofing for rate limiting,
+	// audit logs, and CSRF same-origin comparisons.
+	var trustedProxies []string
+	if s.Dependencies.AppConfig != nil {
+		trustedProxies = s.Dependencies.AppConfig.TrustedProxies
+	}
+	if err := router.SetTrustedProxies(trustedProxies); err != nil {
+		s.Dependencies.Log.Warn().Err(err).Msg("failed to apply trusted proxies; falling back to gin defaults")
+	}
 	router.Use(gin.Recovery())
 
 	router.Use(s.Dependencies.HTTPProvider.SecurityHeadersMiddleware())
diff --git a/internal/server/server.go b/internal/server/server.go
index 2b146486..a98e0534 100644
--- a/internal/server/server.go
+++ b/internal/server/server.go
@@ -10,6 +10,7 @@ import (
 	"github.com/prometheus/client_golang/prometheus/promhttp"
 	"github.com/rs/zerolog"
 
+	"github.com/authorizerdev/authorizer/internal/config"
 	"github.com/authorizerdev/authorizer/internal/graphql"
 	"github.com/authorizerdev/authorizer/internal/http_handlers"
 )
@@ -29,6 +30,7 @@ type Config struct {
 // Dependencies for a server
 type Dependencies struct {
 	Log             *zerolog.Logger
+	AppConfig       *config.Config
 	GraphQLProvider graphql.Provider
 	HTTPProvider    http_handlers.Provider
 }